I regularly have multiple commits being worked on at a time across different parts of the codebase. If I have to sync to head (or any rebase) and one of my side branches that I'm not actively working on hits a merge conflict, I don't have to deal with it in that moment and get distracted from my work at hand (ie: I don't need to context switch). This is a big productivity win for me.
If you want some other points, check out: https://fallthrough.transistor.fm/43#t=0h31m5s
Some points from the episode:
* With no separate index vs commit, (everything is just a commit), you don't need different commands and flags to deal with the different concepts, they are all just marked together. In JJ, if you want to stack/stage something, it's just a normal commit (no reason to have different concepts here).
* You don't have to name/commit a change at all. Every time you run any JJ command (like `jj log`, or `jj status`), it will snapshot the changes you have. This means that if you want to go work on some other branch, you don't have to go and commit your changes (they auto-commit, and you don't have to write a description immediately), then update to master branch and start working again.
* Or you can just `jj split` (https://jj-vcs.github.io/jj/latest/cli-reference/#jj-split), and split a working changeset into 2 separate commits.
I also really like that every change is automatically committed. It’s a great mental model once you get used to it.
Rebases must be done linearly. And right now! Oops, you made an error in an earlier stage of the rebase? Start over, good luck! Want to check something from earlier while you’re in the middle? Sorry, you’re in a modal state and you don’t get to use your regular git tooling.
You can use all the regular git tools in a rebase, in fact it would be quite useless without. You can also just jump to other branches or record a fix to a previous commit. It doesn't matter what you do in the meantime, it only cares what is the HEAD, when you call git rebase --continue, and then it only performs what commands you specify in the rebase todo. You can even change the todo list at any time.
1. We're rewriting some commits. Let's say a chain of commits A through G. We want to make some change to commits A and D.
2. As we're editing commit D, we realized that we need to make some changes to B to match the updated A.
3. Also while editing D we realized that we want to take a look at the state in A to see how something worked there.
With jj, here's what I would do:
1. Run `jj new A`, make the changes, then `jj squash` to squash the changes in to A and propagate them through the chain.
2. Run `jj new D` to make changes. We now notice that we wanted some changes to go into B. We can make the changes in the working copy and run `jj squash --into B -i` to interactively select the changes to squash into B.
3. Run `jj new A` to create a new working-copy commit on top of A, look around in your IDE or whatever, then run `jj edit <old commit on top of D>`. Then run `jj squash` to squash the remaining changes in the working copy into D.
I think you know the steps to do with Git so I won't bother writing them down. I find them less intuitive anyway.
2. Edit changes for D. Also edit changes for B. Interactively `git add` changes for B. `git commit --fixup=B`.
3. Finish changes for D, then call git commit --fixup=D
Before push I run `git rebase --autosquash`, optionally `git rebase --exec='make check'` to rerun all the tests on the changed commits.
Where is that "much harder"?
I wouldn't say I avoid this, I also run `git rebase -i` several times per day, and I also often use `git commit --fixup` during a rebase.
> Makes sense, but doesn't it break down when there are conflicts between the changes you're making in the working copy and the target commit?
Yes, but wouldn't this be the same in JJ, when you do your changes on top of A, and later squash them into D? If you don't want to have the changes, you can also checkout D and do the changes there. Then you have two options:
- `git commit --fixup`, later do `git rebase`
- `git commit --amend`, and `git rebase --onto`
Most times I do the thing described earlier and just solve the conflicts, because that's just a single command. Also when its only a single case, I use `git stash`. (The workflow then is do random fix, git stash, then figure out where these should go, git rebase)
> How do you adjust the steps if there were conflicts between the changes we wanted to make to A and the changes already present in B?
I just resolve them? I think I don't understand this question.
In order to make changes to commit A when there are conflicting changes in B, I was thinking that you would have to use interactive rebase instead because you can no longer make those changes in the working copy and use `git commit --fixup`, right? And because there will now be conflicts in commit B, you will be in this "interrupted rebase" state where you have conflicts in the staging area and it's a bit tricky (IMO) to leave those and look around somewhere else and then come back and resolve the conflicts and continue the rebase later.
> Yes, but wouldn't this be the same in JJ, when you do your changes on top of A, and later squash them into D?
The difference is that we don't end up in an interrupted rebase. If we squashed some changes into A and that resulted in conflicts in B, then we would then create a new working-copy commit on top of the conflicted B (I call all of the related commits B even if they've been rewritten - I hope that's not too confusing). We then resolve the conflicts and squash the resolution into B and the resolution gets propagated to the descendants. We are free at any time to check out any other commit etc.; there's no interrupted rebase or unfinished conflicts we need to take care of first. I hope that clarifies.
Yes.
I don't see the drawback honestly. Invoking git rebase, means I want to resolve conflicts now, when I want to do that later, I can just call git rebase later. When you want to work on top of the B with conflicts, the code wouldn't even compile, so I expect JJ, to just give you the code before the squash, right? How is this different from in Git?
Anyway, I'm just explaining how jj works and what I prefer. As Steve always says, you should use the tools you prefer :)
Yes, I do the same. I think it's not too different. You can also commit randomly somewhere else, it is only a problem once you try to start another rebase or merge. (But I never needed to do it, I just tried it out during discussions like this.)
> Anyway, I'm just explaining how jj works and what I prefer. As Steve always says, you should use the tools you prefer :)
Sure. I'm not objecting to you using JJ, I was objecting to you stating, that it is "much harder" in Git. This is seems to be a common sentiment among JJ users, but it always seem to amount to that people bother to read the manual and understand the tool AFTER they used a VCS for years.
Perhaps, but I don't think that's true for me (or for Steve). I've contributed something like 90 patches to Git itself (mostly to the rebase code). To be fair, that was a while ago.
My impression is actually that many people who disagree with the sentiment that jj is much easier to use seem to have not read its manual :) Some of them seem to have not even tried it. So, the way it looks to me, it's usually the people who argue for jj who have a better understanding of the differences between the two tools.
Have you tried jj yourself and/or have you read some of the documentation?
So actually no, and you have a point. :-)
I often just read "this is hard in Git" and think isn't this just this command? JJ has some nice features, but what appeals to me seems to not to be that hard to add to Git, so I will just wait a bit.
A few months back I had to sanitize the commit history of a repo that needed certain files removed from the entire history. I could have read the manpage for `git filter-branch`, but in jj editing commits is just a normal, sane part of your workflow. It was a blindingly obvious one-liner that just used all the commands I use every day already.
Even better, it was fast. Because “edit the contents of a bunch of commits” is a first-class operation in jj, there’s no need to repeatedly check out and re-commit. Everything can be done directly against the backing repository.
I don't consider `commit --fixup` to be some arcane workaround, that is basically the default to record a change to some older commit.
Editing commits is also a normal, sane part of my workflow, what else is a version control system supposed to do? I consider modifying every commit in a repo not to be that frequent, but nice if JJ supports that easily. Do you want to educate us of the command?
Git also does certain modifications entirely in memory, but when I edit some file obviously my editor needs to access it. Also I want to rerun the tests on some modified commit anyway, so to me checking it out is not some extra cost.
Not sure that they had in mind but you can do `jj squash --from <oldest commit with unwanted file>:: --destination 'root()' <path to unwanted file>`. That will take the changes to the unwanted file from all those commits and move them into a new commit based on the root commit (the root commit is virtual commit that's the ancestor of every other commit).
feels the same when people say, "jq is hard i use python instead" like ok
Two examples from recent memory:
Someone merged the develop branch into their branch, then changed their mind and reverted the merge commit specifically (i.e. reversing all the incoming changes), then somehow merged all of this into the develop branch, undoing weeks of work without noticing. I had to go in and revert the revert to undo the mistake. Yes they messed up, but these things happen with enough people and time.
Another very interesting issue that happened to a less technical person on the team was that their git UI somehow opened the terminal in the wrong folder. They then tried to run some command which made git suggest to run 'git init', creating another git repo in that wrong location. Fast forward some days and we had an issue where people had to clean their repos, so I was in a call with the person helping them run the clean command. The UI opened the wrong location again, I helped them put in the command and it started cleaning. The problem was that this git repo was essentially at the top level on their disk, and since was a fresh repo every single file was considered a new file so it tried to delete EVERYTHING on their disk. This was of course mostly my fault for not running git status before the clean command, but this potential scenario was definitely not on my mind.
An example of something that the reflog isn't going to capture is a git reset --hard losing your unstaged changes, whereas the equivalent flow and commands in jj would allow you to get those contents back.
Git does version everything that is in the repository and all these states occur in the reflog.
Yes. I think that this difference is what introduces a lot of friction, both in the model, and how people use it. The divergence between the files that exist on disk inside your working copy and what's actually tracked means lots of opportunities for friction that go away once you decide that it should. That doesn't mean things are perfect, for example, by default jj only snapshots the filesystem when you run a `jj` command, so you can still lose changes from in between those, you need to enable Watchman to get truly full logging here.
> all these states occur in the reflog.
Well, let's go back to the documentation for reflog:
> Reference logs, or "reflogs", record when the tips of branches and other references were updated in the local repository.
It only tracks changes to refs. That is, the states that refs have been in. So, one big example is detatched HEADs: any changes you make to those, which still are contents of the repository, are not tracked in the reflog.
Even for refs, there's differences: the reflog says "ref was in state x and changed to state y" without any more details. jj's oplog keeps track of not only the state change, but the reason why: "rebased commit <sha> with these args: jj rebase -r <sha> -d trunk"
The reflog only tracks individual refs. Say we rebase multiple commits. The reflog still just says "the head of this branch was in state x and changed to state y" but the oplog says "a rebase happened, it affected all of these commits refs in these ways," that is, it's just inherently more rich in what it tracks, and does it across all relative commits, not only the refs.
This doesn't mean the reflog is bad! It's just a very specific thing. Git could have an operation log too, it's just a different feature.
$ git checkout HEAD
$ git commit --allow-empty -m "_"
$ git checkout master
$ git reflog
a91 (HEAD -> master, origin/master, origin/HEAD) HEAD@{0}: checkout: moving from b94 to master
b94 HEAD@{1}: commit: _
28d (origin/feature, feature) HEAD@{2}: checkout: moving from feature to @
> Even for refs, there's differences: the reflog says "ref was in state x and changed to state y" without any more details. jj's oplog keeps track of not only the state change, but the reason why: "rebased commit <sha> with these args: jj rebase -r <sha> -d trunk"> The reflog only tracks individual refs. Say we rebase multiple commits. The reflog still just says "the head of this branch was in state x and changed to state y" but the oplog says "a rebase happened, it affected all of these commits refs in these ways," that is, it's just inherently more rich in what it tracks, and does it across all relative commits, not only the refs.
68e HEAD@{15}: rebase (finish): returning to refs/heads/feature
68e HEAD@{16}: rebase (pick): message #6
7ff HEAD@{17}: rebase (pick): message #5
797 HEAD@{18}: rebase (pick): message #4
776 HEAD@{19}: rebase (pick): message #3
c7d HEAD@{20}: rebase (pick): message #2
f10 HEAD@{21}: rebase (pick): message #1
c0d HEAD@{22}: rebase (start): checkout @~6
a7c HEAD@{100}: rebase (reword): message ...
3b1 HEAD@{229}: rebase (reset): '3b1' message ...
4a4 HEAD@{270}: rebase (continue): message ...jj still ends up keeping information in here that the reflog doesnt, but you're right that these aren't the strongest points.
--update-refs, --no-update-refs
Automatically force-update any branches that point to commits that
are being rebased. Any branches that are checked out in a worktree
are not updated in this way.
If the configuration variable rebase.updateRefs is set, then this
option can be used to override and disable this setting.Yes, this is something, that JJ provides and Git does not.
I think Git will just gain a oplog. You just need to append a commit hash to a list before each command and implement undo as remove item and checkout. The hardest thing will be race conditions, but Git already knows how to be a database.
Git has lots of sharp edges that can get hairy or at least tedious really rapidly. You have to keep a ton of random arcana in working memory at all times. And a bunch of really useful, lovely workflows are so much of a pain in the ass that you don’t even conceive of doing them.
I learned jj in one day and never went back.
I'm not a welder or a metalworker, but I do know how to weld. I use a welder a handful of times per year when I need/want to. Welding is dangerous, and achieving excellence is a difficult and long road. But I can use the same tools as a pro and still get a few pieces of metal stuck together without having to relearn and restudy the whole system each time something goes wrong.
I haven't used jj in anger yet, but I think it might at least be approaching that style of developer tool.
You can see where you were and hard reset back, no matter what state you are in.
Now, once you teach them it’s a commit graph with names, some of them floating, some people get it.
The thing is, not everyone is comfortable with a commit graph, and most people are not - just like people get lists and arrays but graphs are different.
So I agree with you on principle ( it shouldn’t be difficult), but most people don’t have a graph as a mental model of anything, and I think that’s the biggest obstacle.
The feeling of complexity comes from not yet understanding that commits are just sets of changes to files. They are then thrown off the scent by new terms like origin clone vs push and pull, merge vs rebase, HEAD increment notation vs other commit hashes.
Once people start with a local understanding of using git diff and git add -p they usually get the epiphany. Then git rebase -i and git reflog take them the rest of the way. Then add the distributed push and fetch/pull concepts.
I don't see this anywhere nearly as drastically with other tools.
It’s exactly like folks in 1995 telling you to rtfm when you’re trying to install Linux from a floppy disk. It’s doable, but annoying, and it’s not that easy.
Just in case, I'm talking about the Pro Git book [0]. I remember reading it on my kindle while commuting to office by train. It was so easy to understand, I didn't even need a computer to try things. And it covers everything from bare basics, to advanced topics that get you covered (or at least give you a good head start) if you decide to develop your own jujutsu or kurutu or whutuvur.
The book says that ‘ To really understand the way Git does branching, we need to take a step back and examine how Git stores its data’ then it starts talking about trees and blobs.
At that point you’ve lost almost everyone. If you have a strong interest in vcs implementation then fine, otherwise it’s typically the kind of details you don’t want to hear about. Most people won’t read an entire book just to use a vcs, when what they actually want to hear is ‘this is a commit graph with pointers’.
I agree with you : the information is there. However I don’t think you can in good faith tell most people to rtfm this, and that was my point.
I can, and I do.
The explanation in that book creates a strong coherent and simple mental model of git branching. I honestly can't think of a better explanation. Shorter? Maybe. But "graph with pointers" wouldn't explain it.
So let's agree to disagree.
The git mental model is more complex than cvs, but strangely enough the docs almost invariably refer to the internal implementation details which shouldn’t be needed to work with it.
I remember when git appeared - the internet was full of guides called ‘git finally explained ‘ , and they all started by explaining the plumbing and the implementation. I think this has stuck, and does not make things easy to understand.
Please note I say all this having been using git for close to 20 years, being familiar with the git codebase, and understanding it very well.
I just think the documentation and ui work very hard towards making it difficult to understand .
Hell, I've personally mentored people who struggled with git and I could feel their struggle.
I'm not saying that learning git was an insurmountable task for them, but their struggle was not something that I had to go through.
But whatever state I'm in, I'm sure I can reset back to where I were using reflog.
It does `git pull <default remote> foo`.
> vs a remote
It gives an error, because you haven't specified the remote.
I don't know what behaviour you find intuitive here?
> how do you fix it if you messed up/which is preferred when both exist?
git pull is YOLO mode, so I never do it, but I would just reset the branches where I want them to be? You get a summary with the old and new commit hashes, so resetting is really easy.
And the significant time and effort amounts to a total of 15 seconds.
I understand how to do these things, but git’s interaction model makes it tedious at best and hard at worst.
jj’s interaction model makes these things simple, straightforward, and obvious in the overwhelming majority of cases.
You can do this, but due to file locking, you can corrupt the state if it's shared. jj is specifically designed so that it won't corrupt the repo in this way: https://jj-vcs.github.io/jj/latest/technical/concurrency/
That said, I haven't tried this lately, maybe it's gotten more robust over time. But historically, even a bare repo on something like Dropbox has issues.
I fully agree that this is a niche use-case.
This wasn't what I was talking about, I meant that you should create a bare repo and push to it, not that you work directly in a directory in Dropbox.
This is a very git command.
It calls "git" to "init"ialize a repository, which we don't need a working tree for ("bare") and that it's going to be "shared" with members of the "group".
Here: git rebase is slightly broken in conflict handling. It can be made simpler to understand with jj.
You sure that exists?
Git repositories that contain terabytes of source code?
I could imagine a repo that is terabytes but has binaries committed or similar... But source code?
Having 86 TB of plain text/source code - I can't fathom the scale, honestly
To find bugs, I use "bisect but visually" (I usually use jetbrains IDEs, so I just go to the git history, and do binary search in the commits, displaying all the files that were affected, and jumping easily to such versions).
Git conflicts are easily solvable as well with a gui (jetbrain IDEs) via the cli or via something like Sourcetree. Easily, the most used "feature" of git that i use is:
- for a given line of code, see all the files that were touched when that line was introduced
But I usually do that via the IDE (because to go through dozens of files via cli is a bit of a hassle for me)
So, what am I missing? I know jujutsu is much simple (and powerful) than git, but I only have used the "good parts" of git and it has never been a bottleneck... but ofc, you don't know what you don't know.